[johnmatthewtennant]

Summary

This fixes a bug: when a security alert has no detector-written explanation, Berd should generate one, but it can fail with "An explanation could not be generated" and nothing is logged.

inferSecurityExplanation runs the explanation in a new hidden session, and that session inherits the user's default thinking effort. At high or max effort, the model can spend the whole 20 s budget thinking before it writes any text. promptForText then returns null, and the catch discards every other error, so the cause never reaches a log.

This PR:

• sets thinking_effort to low on the hidden session after setModel (which can rebuild that option) and before the prompt. Not off: always-on thinking models such as Fable 5 treat off as high, so it would raise effort for a user whose default is low. OpenAI reasoning models already send low for off. A model without the option keeps its default, and the explanation still runs.
• logs the failure reason to the app log through logRendererEvent: security explanation failed: no reply text within 20000 ms for a timeout or empty reply, or the formatted ACP error.

Related issue

None found.

Testing

Red/green through the production path: Berd's inferSecurityExplanation ran against the pinned Goose goose serve over ACP. The model was a local stub that records the effort Goose sends and takes longer than the 20 s budget at high, so model latency was simulated.

Model	User default	Requested by Berd	Effort sent	Result
Fable 5	high	nothing (main)	high	no explanation (20 s timeout)
Fable 5	low	off	high	no explanation (20 s timeout)
Fable 5	high or low	low (this PR)	low	explanation returned
Opus 5	high	low (this PR)	low (adaptive)	explanation returned
OpenAI o3	high	nothing (main)	high	no explanation (20 s timeout)
OpenAI o3	high	low (this PR)	low	explanation returned

Live calls with a user default of high, two per model and effort, all returned an explanation inside the budget. At low, first text arrived in 1.8 to 2.8 s on Opus 5 (6.8 to 8.7 s at inherited high) and 1.8 to 1.9 s on Fable 5 (2.0 to 5.9 s).

Unit tests in inferExplanation.test.ts cover the fixed contract:

• effort low is set between setModel and the prompt (fails on main and with off);
• both failure reasons are logged (both fail on main);
• an explanation still returns when setting the option fails (a regression guard).

Not verified: opening an alert without a detector-written explanation in a packaged build. The runs above call inferSecurityExplanation directly rather than going through the alert UI.

🤖 Generated with Claude Code